Atlas

Models

← All models

Grok 3

xAI · 2025-02-19

Grok 3 is xAI's flagship model released in February 2025, trained on the Colossus supercluster with roughly 10× the compute of its predecessor and featuring large-scale reinforcement learning for extended chain-of-thought reasoning. It introduced a Think mode and Big Brain mode for multi-step problem solving, a built-in DeepSearch capability for real-time web retrieval, and a 1 million-token context window. Grok 3 demonstrated significant improvements over Grok 2 in mathematics, coding, world knowledge, and instruction following.

Benchmark scores

BenchmarkScore
A-Fantasia47.0
A-Fantasia Backwards Spelling36.0
A-Fantasia Chess28.0
A-Fantasia Cube Rotation77.0
AA-LCR54.7
AA-Omniscience Index-32.3
AidanBench1796
Aider Polyglot53.3
AIME 2024 (avg@8)60.4
AIME 2024-202558.8
AIME 202577.3
AIME 2025 (avg@8)57.1
APEX-Agents2.1
ARC-AGI-15.5
ARC-AGI-20.0
Artificial Analysis Intelligence Index18.4
Artificial Analysis Omniscience Accuracy28.3
Artificial Analysis Omniscience Hallucination Rate84.6
BALROG29.5
Capability140.8
Confabulations Leaderboard Confabulation Rate17.8
Confabulations Leaderboard Non-Response Rate10.6
Confabulations Leaderboard Weighted Score14.2
CorpFin v259.7
CorpFin v2 - Exact Pages58.5
CorpFin v2 - Max Fitting Context57.1
CorpFin v2 - Shared Max Context63.5
Creative Writing v3 Elo1180
Creative Writing v3 Rubric Score15.4
CritPt0.0
Loading Atlas data…